Broadcast news LM adaptation over time

On the dynamic adaptation of language models based on dialogue information

“…This approach is known as count merging (Federico & Bertoldi, 2004;Ljolje, Hindle, Riley, & Sproat, 2000;Lobacheva, 2000), and it is usually related to well-known adaptation approaches, such as Maximum a Posteriori (MAP) or Maximum Likelihood Linear Regression (MLLR). MAP has been successfully applied to LM adaptation (see, for instance, (Chen et al, 2001;Liu et al, 2008)), although it has been proven in Bacchiani, Riley, Roark, and Sproat (2006), Bacchiani and Roark (2003), and Hsu (2007) that the performance of a MAP-based adaptation system is similar to that achieved with linear interpolation, but requiring a greater computational effort.…”

“…One of the most common sources of data for adaptation purposes is the Internet. It is usually queried when trying to adapt the LMs to a specific topic (Lecorvé, Gravier, & Sébillot, 2009;Shi et al, 2008), or to the most recent content, such as in the case of a Broadcast News transcription domain, as proposed in Federico and Bertoldi (2004), Martins et al (2010), Saykham, Chotimongkol, and Wutiwiwatchai (2010).…”

Dynamic language modeling for European Portuguese

“…In [Federico and Bertoldi, 2004], an approach of this kind was introduced. The baseline vocabulary of 62K words is extended by adding to that special OOV word class 60K new words selected from the contemporary written news on a daily basis.…”

“…In [Federico and Bertoldi, 2004] the problem of updating over time the LM component of an Italian broadcast news transcription system is addressed. In particular, vocabulary adaptation, done on a daily basis, is carried out by adding words to the active vocabulary according to frequency and recency in contemporary written news, which allowed achieving significantly lower OOV word rates.…”

Dynamic Vocabulary Adaptation for a daily and real-time Broadcast News Transcription System

“…The idea of vocabulary and LM adaptation is to use written news daily available on the Internet to adjust the vocabulary and reduce the impact of linguistic differences over time. Similar approaches to the one presented here have been proposed where vocabulary adaptation is carried out by adding and removing words from the baseline vocabulary according to frequency and recency in contemporary written news [2] [3]. In [1] a different approach is suggested which uses a multi-pass recognition strategy to generate morphological variations of the list of all words in the lattice, thus dynamically adapting the recognition vocabulary.…”